A SHORT PROOF OF THE COIFMAN-MEYER MULTILINEAR

THEOREM

CAMIL MUSCALU, JILL PIPHER, TERENCE TAO, AND CHRISTOPH THIELE

Abstract. We give a short proof of the well known Coifman-Meyer theorem on multi-
linear operators.

1. Introduction
The main task of the present paper is to present a new proof of the classical Coifman-
Meyer theorem on multilinear singular integrals, see [2], [3], [5], [6].
Let m L (IRn ) be a bounded function which is smooth away from the origin and
satisfies the following Marcinkiewicz-Mihlin-Hormander type condition

1
| m()| . , (1)
||||
for sufficiently many multiindices 1. For f1 , ..., fn S(IR) Schwartz functions on the
real line, we define the n-linear operator Tm by the formula

Z
Tm (f1 , ..., fn )(x) = n
m()fb1 (1 )...fbn (n )e2ix(1 +...+n ) d1 ...dn . (2)
IR
The following theorem holds, see [2], [3], [5], [6].
Theorem 1.1. As defined, the multilinear operator Tm maps Lp1 ... Lpn Lp as long
as 1 < pi , 1 i n, 1/p1 + ... + 1/pn = 1/p and 0 < p < .
When such an n + 1-tuple (p1 , ..., pn , p), has the property that 0 < p < 1 and pj =
for some 1 j n then, for some technical reasons (see [5], [7]), by L one actually
means L c the space of bounded measurable functions with compact support.
The case p 1 has been proven in [3] while the general case p > 1/n has been
independently settled in [6] and [5]. The interesting fact that p can be a number smaller
than 1 goes back to [2]. The usual argument to prove the theorem (see [2], [3], [5], [6]) uses
the celebrated T 1 theorem of David and Journe [4] and relies on BM O theory, Carleson
measures and C.Feffermans duality theorem between the Hardy space H 1 and BM O.
As the reader will see, our proof is conceptually simpler and does not use any of the
aforementioned ingredients. It is based on a careful stopping time argument involving the
Hardy-Littlewood maximal function and the Littlewood-Paley square function.

1Throughout the paper we will write A . B iff there is a universal constant C > 0 so that A CB.
1
2 CAMIL MUSCALU, JILL PIPHER, TERENCE TAO, AND CHRISTOPH THIELE

2. Model operators
For simplicity we treat the n = 2 case only. It is a standard fact by now (see for
instance the papers [7], [8]) that the study of our bilinear operators Tm can be reduced
to the study of finitely many discrete model operators jP , j = 0, 1, 2, 3 of the form

X 1
jP (f1 , f2 ) = P hf1 , P1 ihf2 , P2 iP3 . (3)
P P
|IP |1/2

Here P is a collection of lacunary dyadic tiles corresponding to lattice points (k, n) in Z 2

and (P )P is an arbitrary sequence of uniformly bounded constants. More precisely, Pi
i = 1, 2, 3 are defined by Pi = IP Pi where IP = [n2k , (n + 1)2k+1 ], Pi = [0, 2k ] if
i = j and Pi = [2k , 2k+1 ] if i 6= j, while Pi i = 1, 2, 3 are L2 normalized wave packets
corresponding to the Heisenberg boxes Pi i = 1, 2, 3. This means that the function Pi is a
smooth L2 normalized bump function adapted to the interval IP whose Fourier transform
cPi is supported inside the interval Pi for i = 1, 2, 3. We should emphasize here that the
tile P is uniquely determined by the interval IP .
To explain this reduction in a few words, let Q := I J be a dyadic rectangle in
the plane, having the property that diam(Q) dist(Q, {0}). Let also I , J be two L1
normalized smooth functions such that supp(bI ) I and supp( cJ ) J. If we replace
the symbol m(1 , 2 ) by bI (1 )
cJ (2 ) we observe that the right hand side of (2) becomes
(f1 I )(x)(f2 J )(x). On the other hand, inequality (1) implies that one can think of
m as being essentially constant on each such Q and so the integral in (2) when smoothly
restricted to this cube, becomes roughly

Q (f1 I )(x)(f2 J )(x).

Then, one covers the plane by a collection of carefully selected such Whitney cubes and
discretize again in the x-variable. In the end, one obtains a formula in which the general
Tm is written as an average of operators of the type of the model operators above. The
details can be found in [7], [8]. Now our analysis of the operators jP will be independent
on j = 0, 1, 2, 3 and so we can assume without loss of generality that j = 1 and write
from now on, for simplicity, P instead of 1P (the reader will observe that for j 6= 1, the
only difference is that the roles of the Hardy Littlewood maximal function M and the
Littlewood Paley square function S, get permuted). It is therefore enough to prove the
theorem for the bilinear operator P .

3. The proof
First, let us observe that it is very easy to obtain the necessary Lp estimates in the
particular case when all the indices are strictly between 1 and . To see this, let f Lp ,
g Lq , h Lr for 1 < p, q, r < with 1/p + 1/q + 1/r = 1. Then,
Z 
  X 1
 P (f, g)(x)h(x) dx . |hf, P1 i||hg, P2 i||hh, P3 i| =
 |I P|
1/2
IR P P
 A SHORT PROOF OF THE COIFMAN-MEYER MULTILINEAR THEOREM 3

Z X |hf, P i| |hg, P i| |hh, P i|

1 2 3
1/2 1/2 1/2
IP (x) dx . (4)
IR P P |I P | |I P | |I P |

Z   X !1/2 !1/2
|hf, P1 i| |hg, P2 i|2 X |hh, P i|2
3
sup 1/2
IP (x) IP (x) IP (x) dx .
IR P P |I P | P P
|I P | P P
|I P |
Z
M f (x)Sg(x)Sh(x) dx . kM f kp kSgkq kShkr .
IR
kf kp kgkq khkr ,
where M is the maximal function of Hardy and Littlewood and S is the discrete square
function of Littlewood and Paley, see [9] and [7]. This means that theorem 1.1 is nontrivial
only when one index is , or less or equal than 1. To prove the general case we just need to
show that the bilinear operator P maps L1 L1 L1/2, because then, by interpolation
and symmetry the theorem follows as in [7]. Let f, g L1 be such that kf k1 = kgk1 = 1.
We now recall Lemma 5.4 in [1].
Lemma 3.1. Let 0 < p < and A > 0. Then the following statements are equivalent
up to constants:
(i) kf kp, . A.
(ii) For every set E with 0 < |E| < , there exists a subset E 0 E with |E 0 | |E| and
0
|hf, E 0 i| . A|E|1/p . Here p0 is defined by 1/p0 + 1/p = 1 (note that p0 can be a negative
number!).
Proof To see that (i) implies (ii), set
E 0 := E \ {x : |f (x)| CA|E|1/p }.
If C is a sufficiently large constant, then (i) implies |E 0 | |E| and the claim follows.
To see that (ii) implies (i), let > 0 be arbitrary and set E := {x : Re(f (x)) > }.
Then by (ii) we have

|E| |E 0 | . A|E|1/p ,
0

and (i) easily follows (replacing Re by Re, Im, Im as necessary).

Using this Lemma 3.1 in the particular case p = 1/2 and the scale invariance, it is
enough to show that given E IR |E| = 1, one can find a subset E 0 E with |E 0 | 1
such that

X 1
1/2
|hf, P1 i||hg, P2 i||hh, P3 i| . 1 (5)
P P
|I P|

where h := E 0 . Fix such a set E with |E| = 1. To construct the subset E 0 , we first
consider
0 = {x IR : M (f )(x) > C} {x IR : S(g)(x) > C} {x IR : M (g)(x) > C}.
Also, define
4 CAMIL MUSCALU, JILL PIPHER, TERENCE TAO, AND CHRISTOPH THIELE

1
= {x IR : M (10 )(x) > }.
100
Clearly, we have || < 1/2, if C is a big enough constant which we fix from now on.
Then, we define E 0 := E \ = E c and observe that indeed |E 0 | 1. After this, we
split our sum in (5) into two parts
X X X
= + := I + II.
P P IP c 6= IP c =
We also assume that the set P is finite, since our estimates do not depend on its cardinality.
First, we estimate term I. Since IP c 6= , it follows that |IP|I
P|
0| 1
< 100 or equivalently,
c 99
|IP 0 | > 100 |IP |.
We are now going to describe three decomposition procedures, one for each function
f, g, h. Later on, we will combine them, in order to handle our sum. First, define

C
1 = {x IR : M (f )(x) > }
21
and set

1
T1 = {P P : |IP 1 | > |IP |},
100
then define

C
2 = {x IR : M (f )(x) > }
22
and set

1
T2 = {P P \ T1 : |IP 2 | > |IP |},
100
and so on. (The constant C > 0 is the one which we fixed before). Since there are finitely
many tiles, this algorithm ends after a while, producing the sets {n } and {Tn } such that
P = n Tn . Independently, define

C
01 = {x IR : S(g)(x) > }
21
and set

1
T01 = {P P : |IP 01 | > |IP |},
100
then define

C
02 = {x IR : S(g)(x) > }
22
and set

1
T02 = {P P \ T01 : |IP 02 | > |IP |},
100
 A SHORT PROOF OF THE COIFMAN-MEYER MULTILINEAR THEOREM 5

and so on, producing the sets {0n } and {T0n } such that P = n T0n . We would like to
have such a decomposition available for the function h also. To do this, we first need to
construct the analogue of the set 0 , for it. We will therefore pick N > 0 a big enough
00 c 99
integer such that for every P P we have |IP N | > 100 |IP | where we defined

00N = {x IR : S(h)(x) > C2N }.

Then, similarly to the previous algorithms, we define

C2N
00N +1 = {x IR : S(h)(x) > }
21
and set
1
T00N +1 = {P P : |IP 00N +1 | > |IP |},
100
then define

C2N
00N +2 = {x IR : S(h)(x) > }
22
and set
1
T00N +2 = {P P \ T00N +1 : |IP 00N +2 | >
|IP |},
100
and so on, constructing the sets {00n } and {T00n } such that P = n T00n . Then we write the
term I as

X X 1
|hf, P1 i||hg, P2 i||hh, P3 i||IP |, (6)
n1 ,n2 >0,n3 >N P Tn1 ,n2 ,n3
|IP |3/2

where Tn1 ,n2 ,n3 := Tn1 T0n2 T00n3 . Now, if P belongs to Tn1 ,n2 ,n3 this means in particular
that P has not been selected at the previous n1 1, n2 1 and n3 1 steps respectively,
1 1 1
which means that |IP n1 1 | < 100 |IP |, |IP 0n2 1 | < 100 |IP | and |IP 00n3 1 | < 100 |IP |
c 99 0 c 99 00 c 99
or equivalently, |IP n1 1 | > 100 |IP |, |IP n2 1 | > 100 |IP | and |IP n3 1 | > 100 |IP |.
But this implies that

0 00 97
|IP cn1 1 nc2 1 nc3 1 | > |IP |. (7)
100
In particular, using (7), the term in (6) is smaller than

X X 1 0 00

3/2
|hf, P1 i||hg, P2 i||hh, P3 i||IP cn1 1 nc2 1 nc3 1 | =
n1 ,n2 >0,n3 >N P Tn1 ,n2 ,n3
|IP |

X Z X 1
|hf, P1 i||hg, P2 i||hh, P3 i|IP (x) dx
n1 ,n2 >0,n3 >N
0 00
cn 1 nc 1 n c 1 P Tn1 ,n2 ,n3
|IP |3/2
1 2 3

X Z
. M (f )(x)S(g)(x)S(h)(x) dx
0 00
n1 ,n2 >0,n3 >N cn nc 1 n c 1 Tn ,n ,n
1 1 2 3 1 2 3
6 CAMIL MUSCALU, JILL PIPHER, TERENCE TAO, AND CHRISTOPH THIELE

X
. 2n1 2n2 2n3 |Tn1 ,n2 ,n3 |, (8)
n1 ,n2 >0,n3 >N

where
[
Tn1 ,n2 ,n3 := IP .
P Tn1 ,n2 ,n3

On the other hand we can write

1
|Tn1 ,n2 ,n3 | |Tn1 | |{x IR : M (n1 )(x) > }|
100
C
. |n1 | = |{x IR : M (f )(x) > }| . 2n1 .
2 n1
Similarly, we have |Tn1 ,n2 ,n3 | . 2n2 and also |Tn1 ,n2 ,n3 | . 2n2 , for every 1, since
|E 0 | 1. In particular, it follows that

|Tn1 ,n2 ,n3 | . 2n1 1 2n2 2 2n3 3 (9)

for any 0 1 , 2 , 3 < 1, such that 1 + 2 + 3 = 1. Now we split the sum in (8) into
X X
2n1 2n2 2n3 |Tn1 ,n2 ,n3 | + 2n1 2n2 2n3 |Tn1 ,n2 ,n3 |.
n1 ,n2 >0,n3 >0 n1 ,n2 >0,0>n3 >N (10)
To estimate the first term in (10) we use the inequality (9) in the particular case 1 =
2 = 1/2, 3 = 0, while to estimate the second term we use (9) for j , j = 1, 2, 3 such that
1 < 1, 2 < 1 and 3 1 > 0. With these choices, the geometric sums in (10) are finite.
This ends the discussion on I.
Now term II is much simpler, being just an error term. We split
[
P := Pd
d>0
where

dist(IP , c )
Pd := {P P : 2d }
|IP |
and easily observe that
X
|IP | . || 1. (11)
P Pd ;IP

Then, term II is smaller than

X X hf, P1 i hg, P2 i hh, P3 i X X

|IP || || || | . |IP |2d 2d 2Kd . 1,
d>0 P Pd ;IP
|IP |1/2 |IP |1/2 |IP |1/2 d>0 P P ;I d P

for any big number K > 0, and this ends the proof.
 A SHORT PROOF OF THE COIFMAN-MEYER MULTILINEAR THEOREM 7

References
[1] Auscher P., Hofmann S., Muscalu C., Tao T. and Thiele C. Carleson measures, trees, extrapolation
and T b theorems, Publ. Mat., vol. 46, 257-325, [2002].
[2] Coifman R. and Meyer Y., On commutators of singular integrals and bilinear singular integrals,
Trans. Amer. Math. Soc. vol. 212, 315-331, [1975].
[3] Coifman R. and Meyer Y., Ondelettes et operateurs III. Operateurs multilineaires, Hermann, Paris,
[1991].
[4] David G. and Journe J-L., A boundedness criterion for generalized Calderon-Zygmund operators,
Ann. of Math., vol. 120, 371-397, [1984].
[5] Grafakos L. and Torres R., Multilinear Calderon-Zygmund theory, Adv. Math., vol. 165, 124-164,
[2002].
[6] Kenig C. and Stein E., Multilinear estimates and fractional integration, Math. Res. Lett., vol. 6,
1-15, [1999].
[7] Muscalu C., Tao T. and Thiele C., Multilinear operators given by singular multipliers, J. Amer.
Math. Soc. 15, [2002], 469-496.
[8] Muscalu C., Pipher J., Tao T. and Thiele, C., Bi-parameter paraproducts, Preprint, [2003].
[9] Stein, E., Harmonic Analysis: Real Variable Methods, Orthogonality, and Oscillatory Integrals.
Princeton University Press, Princeton, [1993].

Department of Mathematics, Cornell University, Ithaca, NY 14853

E-mail address: camil@math.cornell.edu

Department of Mathematics, Brown University, Providence, RI 02912

E-mail address: jpipher@math.brown.edu

Department of Mathematics, UCLA, Los Angeles, CA 90095

E-mail address: tao@math.ucla.edu

Department of Mathematics, UCLA, Los Angeles, CA 90095

E-mail address: thiele@math.ucla.edu